Skip to content

pt-diskstats

NAME

pt-diskstats - An interactive I/O monitoring tool for GNU/Linux.

SYNOPSIS

Usage

pt-diskstats [OPTIONS] [FILES]

pt-diskstats prints disk I/O statistics for GNU/Linux. It is somewhat similar to iostat, but it is interactive and more detailed. It can analyze samples gathered from another machine.

RISKS

Percona Toolkit is mature, proven in the real world, and well tested, but all database tools can pose a risk to the system and the database server. Before using this tool, please:

  • Read the tool’s documentation

  • Review the tool’s known “BUGS”

  • Test the tool on a non-production server

  • Backup your production server and verify the backups

DESCRIPTION

The pt-diskstats tool is similar to iostat, but has some advantages. It prints read and write statistics separately, and has more columns. It is menu-driven and interactive, with several different ways to aggregate the data. It integrates well with the pt-stalk tool. It also does the “right thing” by default, such as hiding disks that are idle. These properties make it very convenient for quickly drilling down into I/O performance and inspecting disk behavior.

This program works in two modes. The default is to collect samples of /proc/diskstats and print out the formatted statistics at intervals. The other mode is to process a file that contains saved samples of /proc/diskstats; there is a shell script later in this documentation that shows how to collect such a file.

In both cases, the tool is interactively controlled by keystrokes, so you can redisplay and slice the data flexibly and easily. It loops forever, until you exit with the ‘q’ key. If you press the ‘?’ key, you will bring up the interactive help menu that shows which keys control the program.

When the program is gathering samples of /proc/diskstats and refreshing its display, it prints information about the newest sample each time it refreshes. When it is operating on a file of saved samples, it redraws the entire file’s contents every time you change an option.

The program doesn’t print information about every block device on the system. It hides devices that it has never observed to have any activity. You can enable and disable this by pressing the ‘i’ key.

OUTPUT

In the rest of this documentation, we will try to clarify the distinction between block devices (/dev/sda1, for example), which the kernel presents to the application via a filesystem, versus the (usually) physical device underneath the block device, which could be a disk, a RAID controller, and so on. We will sometimes refer to logical I/O operations, which occur at the block device, versus physical I/Os which are performed on the underlying device. When we refer to the queue, we are speaking of the queue associated with the block device, which holds requests until they’re issued to the physical device.

The program’s output looks like the following sample, which is too wide for this manual page, so we have formatted it as several samples with line breaks:

#ts device rd_s rd_avkb rd_mb_s rd_mrg rd_cnc   rd_rt
{6} sda     0.9     4.2     0.0     0%    0.0    17.9
{6} sdb     0.4     4.0     0.0     0%    0.0    26.1
{6} dm-0    0.0     4.0     0.0     0%    0.0    13.5
{6} dm-1    0.8     4.0     0.0     0%    0.0    16.0

    ...    wr_s wr_avkb wr_mb_s wr_mrg wr_cnc   wr_rt
    ...    99.7     6.2     0.6    35%    3.7    23.7
    ...    14.5    15.8     0.2    75%    0.5     9.2
    ...     1.0     4.0     0.0     0%    0.0     2.3
    ...   117.7     4.0     0.5     0%    4.1    35.1

    ...              busy in_prg    io_s  qtime stime
    ...                6%      0   100.6   23.3   0.4
    ...                4%      0    14.9    8.6   0.6
    ...                0%      0     1.1    1.5   1.2
    ...                5%      0   118.5   34.5   0.4

The columns are as follows:

#ts

This column’s contents vary depending on the tool’s aggregation mode. In the default mode, when each line contains information about a single disk but possibly aggregates across several samples from that disk, this column shows the number of samples that were included into the line of output, in {curly braces}. In the example shown, each line of output aggregates {10} samples of /proc/diskstats.

In the “all” group-by mode, this column shows timestamp offsets, relative to the time the tool began aggregating or the timestamp of the previous lines printed, depending on the mode. The output can be confusing to explain, but it’s rather intuitive when you see the lines appearing on your screen periodically.

Similarly, in “sample” group-by mode, the number indicates the total time span that is grouped into each sample.

If you specify --show-timestamps, this field instead shows the timestamp at which the sample was taken; if multiple timestamps are present in a single line of output, then the first timestamp is used.

device

The device name. If there is more than one device, then instead the number of devices aggregated into the line is shown, in {curly braces}.

rd_s

The average number of reads per second. This is the number of I/O requests that were sent to the underlying device. This usually is a smaller number than the number of logical IO requests made by applications. More requests might have been queued to the block device, but some of them usually are merged before being sent to the disk.

This field is computed from the contents of /proc/diskstats as follows. See “KERNEL DOCUMENTATION” below for the meaning of the field numbers:

delta[field1] / delta[time]

rd_avkb

The average size of the reads, in kilobytes. This field is computed as follows:

2 * delta[field3] / delta[field1]

rd_mb_s

The average number of megabytes read per second. Computed as follows:

2 * delta[field3] / delta[time]

rd_mrg

The percentage of read requests that were merged together in the queue scheduler before being sent to the physical device. The field is computed as follows:

100 * delta[field2] / (delta[field2] + delta[field1])

rd_cnc

The average concurrency of the read operations, as computed by Little’s Law. This is the end-to-end concurrency on the block device, not the underlying disk’s concurrency. It includes time spent in the queue. The field is computed as follows:

delta[field4] / delta[time] / 1000 / devices-in-group

rd_rt

The average response time of the read operations, in milliseconds. This is the end-to-end response time, including time spent in the queue. It is the response time that the application making I/O requests sees, not the response time of the physical disk underlying the block device. It is computed as follows:

delta[field4] / (delta[field1] + delta[field2])

wr_s, wr_avkb, wr_mb_s, wr_mrg, wr_cnc, wr_rt

These columns show write activity, and they match the corresponding columns for read activity.

busy

The fraction of wall-clock time that the device had at least one request in progress; this is what iostat calls %util, and indeed it is utilization, depending on how you define utilization, but that is sometimes ambiguous in common parlance. It may also be called the residence time; the time during which at least one request was resident in the system. It is computed as follows:

100 * delta[field10] / (1000 * delta[time])

This field cannot exceed 100% unless there is a rounding error, but it is a common mistake to think that a device that’s busy all the time is saturated. A device such as a RAID volume should support concurrency higher than 1, and solid-state drives can support very high concurrency. Concurrency can grow without bound, and is a more reliable indicator of how loaded the device really is.

in_prg

The number of requests that were in progress. Unlike the read and write concurrencies, which are averages that are generated from reliable numbers, this number is an instantaneous sample, and you can see that it might represent a spike of requests, rather than the true long-term average. If this number is large, it essentially means that the device is heavily loaded. It is computed as follows:

field9

ios_s

The average throughput of the physical device, in I/O operations per second (IOPS). This column shows the total IOPS the underlying device is handling. It is the sum of rd_s and wr_s.

qtime

The average queue time; that is, time a request spends in the device scheduler queue before being sent to the physical device. This is an average over reads and writes.

It is computed in a slightly complex way: the average response time seen by the application, minus the average service time (see the description of the next column). This is derived from the queueing theory formula for response time, R = W + S: response time = queue time + service time. This is solved for W, of course, to give W = R - S. The computation follows:

delta[field11] / (delta[field1, 2, 5, 6] + delta[field9])
   - delta[field10] / delta[field1, 2, 5, 6]

See the description for stime for more details and cautions.

stime

The average service time; that is, the time elapsed while the physical device processes the request, after the request finishes waiting in the queue. This is an average over reads and writes. It is computed from the queueing theory utilization formula, U = SX, solved for S. This means that utilization divided by throughput gives service time:

delta[field10] / (delta[field1, 2, 5, 6])

Note, however, that there can be some kernel bugs that cause field 9 in /proc/diskstats to become negative, and this can cause field 10 to be wrong, thus making the service time computation not wholly trustworthy.

Note that in the above formula we use utilization very specifically. It is a duration, not a percentage.

You can compare the stime and qtime columns to see whether the response time for reads and writes is spent in the queue or on the physical device. However, you cannot see the difference between reads and writes. Changing the block device scheduler algorithm might improve queue time greatly. The default algorithm, cfq, is very bad for servers, and should only be used on laptops and workstations that perform tasks such as working with spreadsheets and surfing the Internet.

If you are used to using iostat, you might wonder where you can find the same information in pt-diskstats. Here are two samples of output from both tools on the same machine at the same time, for /dev/sda, wrapped to fit:

     #ts dev rd_s rd_avkb rd_mb_s rd_mrg rd_cnc   rd_rt
08:50:10 sda  0.0     0.0     0.0     0%    0.0     0.0
08:50:20 sda  0.4     4.0     0.0     0%    0.0    15.5
08:50:30 sda  2.1     4.4     0.0     0%    0.0    21.1
08:50:40 sda  2.4     4.0     0.0     0%    0.0    15.4
08:50:50 sda  0.1     4.0     0.0     0%    0.0    33.0

             wr_s wr_avkb wr_mb_s wr_mrg wr_cnc   wr_rt
              7.7    25.5     0.2    84%    0.0     0.3
             49.6     6.8     0.3    41%    2.4    28.8
            210.1     5.6     1.1    28%    7.4    25.2
            297.1     5.4     1.6    26%   11.4    28.3
             11.9    11.7     0.1    66%    0.2     4.9

                     busy  in_prg   io_s  qtime   stime
                       1%       0    7.7    0.1     0.2
                       6%       0   50.0   28.1     0.7
                      12%       0  212.2   24.8     0.4
                      16%       0  299.5   27.8     0.4
                       1%       0   12.0    4.7     0.3

         Dev rrqm/s  wrqm/s   r/s    w/s  rMB/s  wMB/s
08:50:10 sda   0.00   41.40  0.00   7.70   0.00   0.19
08:50:20 sda   0.00   34.70  0.40  49.60   0.00   0.33
08:50:30 sda   0.00   83.30  2.10 210.10   0.01   1.15
08:50:40 sda   0.00  105.10  2.40 297.90   0.01   1.58
08:50:50 sda   0.00   22.50  0.10  11.10   0.00   0.13

                avgrq-sz avgqu-sz  await  svctm  %util
                   51.01     0.02   2.04   1.25   0.96
                   13.55     2.44  48.76   1.16   5.79
                   11.15     7.45  35.10   0.55  11.76
                   10.81    11.40  37.96   0.53  15.97
                   24.07     0.17  15.60   0.87   0.97

The correspondence between the columns is not one-to-one. In particular:

rrqm/s, wrqm/s

These columns in iostat are replaced by rd_mrg and wr_mrg in pt-diskstats.

avgrq-sz

This column is in sectors in iostat, and is a combination of reads and writes. The pt-diskstats output breaks these out separately and shows them in kB. You can derive it via a weighted average of rd_avkb and wr_avkb in pt-diskstats, and then multiply by 2 to get sectors (each sector is 512 bytes).

avgqu-sz

This column really represents concurrency at the block device scheduler. The pt-diskstats output shows concurrency for reads and writes separately: rd_cnc and wr_cnc.

await

This column is the average response time from the beginning to the end of a request to the block device, including queue time and service time, and is not shown in pt-diskstats. Instead, pt-diskstats shows individual response times at the disk level for reads and writes (rd_rt and wr_rt), as well as queue time versus service time for reads and writes in aggregate.

svctm

This column is the average service time at the disk, and is shown as stime in pt-diskstats.

%util

This column is called busy in pt-diskstats. Utilization is usually defined as the portion of time during which there was at least one active request, not as a percentage, which is why we chose to avoid this confusing term.

COLLECTING DATA

It is straightforward to gather a sample of data for this tool. Files should have this format, with a timestamp line preceding each sample of statistics:

TS <timestamp>
<contents of /proc/diskstats>
TS <timestamp>
<contents of /proc/diskstats>
... et cetera

You can simply use pt-diskstats with --save-samples to collect this data for you. If you wish to capture samples as part of some other tool, and use pt-diskstats to analyze them, you can include a snippet of shell script such as the following:

INTERVAL=1
while true; do
   sleep=$(date +%s.%N | awk "{print $INTERVAL - (\$1 % $INTERVAL)}")
   sleep $sleep
   date +"TS %s.%N %F %T" >> diskstats-samples.txt
   cat /proc/diskstats >> diskstats-samples.txt
done

KERNEL DOCUMENTATION

This documentation supplements the official documentation on the contents of /proc/diskstats. That documentation can sometimes be difficult to understand for those who are not familiar with Linux kernel internals. The contents of /proc/diskstats are generated by the diskstats_show() function in the kernel source file block/genhd.c.

Here is a sample of /proc/diskstats on a recent kernel.

8 1 sda1 426 243 3386 2056 3 0 18 87 0 2135 2142

The fields in this sample are as follows. The first three fields are the major and minor device numbers (8, 1), and the device name (sda1). They are followed by 11 fields of statistics:

  1. The number of reads completed. This is the number of physical reads done by the underlying disk, not the number of reads that applications made from the block device. This means that 426 actual reads have completed successfully to the disk on which /dev/sda1 resides. Reads are not counted until they complete.

  2. The number of reads merged because they were adjacent. In the sample, 243 reads were merged. This means that /dev/sda1 actually received 869 logical reads, but sent only 426 physical reads to the underlying physical device.

  3. The number of sectors read successfully. The 426 physical reads to the disk read 3386 sectors. Sectors are 512 bytes, so a total of about 1.65MB have been read from /dev/sda1.

  4. The number of milliseconds spent reading. This counts only reads that have completed, not reads that are in progress. It counts the time spent from when requests are placed on the queue until they complete, not the time that the underlying disk spends servicing the requests. That is, it measures the total response time seen by applications, not disk response times.

  5. Ditto for field 1, but for writes.

  6. Ditto for field 2, but for writes.

  7. Ditto for field 3, but for writes.

  8. Ditto for field 4, but for writes.

  9. The number of I/Os currently in progress, that is, they’ve been scheduled by the queue scheduler and issued to the disk (submitted to the underlying disk’s queue), but not yet completed. There are bugs in some kernels that cause this number, and thus fields 10 and 11, to be wrong sometimes.

  10. The total number of milliseconds spent doing I/Os. This is not the total response time seen by the applications; it is the total amount of time during which at least one I/O was in progress. If one I/O is issued at time 100, another comes in at 101, and both of them complete at 102, then this field increments by 2, not 3.

  11. This field counts the total response time of all I/Os. In contrast to field 10, it counts double when two I/Os overlap. In our previous example, this field would increment by 3, not 2.

OPTIONS

This tool accepts additional command-line arguments. Refer to the “SYNOPSIS” and usage information for details.

--columns-regex

type: string; default: .

Print columns that match this Perl regex.

--config

type: Array

Read this comma-separated list of config files; if specified, this must be the first option on the command line.

--devices-regex

type: string

Print devices that match this Perl regex.

--group-by

type: string; default: all

Group-by mode: disk, sample, or all. In disk mode, each line of output shows one disk device, with the statistics computed since the tool started. In sample mode, each line of output shows one sample of statistics, with all disks averaged together. In all mode, each line of output shows one sample and one disk device.

--headers

type: Hash; default: group,scroll

If group is present, each sample will be separated by a blank line, unless the sample is only one line. If scroll is present, the tool will print the headers as often as needed to prevent them from scrolling out of view. Note that you can press the space bar, or the enter key, to reprint headers at will.

--help

Show help and exit.

--interval

type: int; default: 1

When in interactive mode, wait N seconds before printing to the screen. Also, how often the tool should sample /proc/diskstats.

The tool attempts to gather statistics exactly on even intervals of clock time. That is, if you specify a 5-second interval, it will try to capture samples at 12:00:00, 12:00:05, and so on; it will not gather at 12:00:01, 12:00:06 and so forth.

This can lead to slightly odd delays in some circumstances, because the tool waits one full cycle before printing out the first set of lines. (Unlike iostat and vmstat, pt-diskstats does not start with a line representing the averages since the computer was booted.) Therefore, the rule has an exception to avoid very long delays. Suppose you specify a 10-second interval, but you start the tool at 12:00:00.01. The tool might wait until 12:00:20 to print its first lines of output, and in the intervening 19.99 seconds, it would appear to do nothing.

To alleviate this, the tool waits until the next even interval of time to gather, unless more than 20% of that interval remains. This means the tool will never wait more than 120% of the sampling interval to produce output, e.g if you start the tool at 12:00:53 with a 10-second sampling interval, then the first sample will be only 7 seconds long, not 10 seconds.

--iterations

type: int

When in interactive mode, stop after N samples. Run forever by default.

--sample-time

type: int; default: 1

In –group-by sample mode, include N seconds of samples per group.

--save-samples

type: string

File to save diskstats samples in; these can be used for later analysis.

--show-inactive

Show inactive devices.

--show-timestamps

Show a ‘HH:MM:SS’ timestamp in the #ts column. If multiple timestamps are aggregated into one line, the first timestamp is shown.

--version

Show version and exit.

--[no]version-check

default: yes

Check for the latest version of Percona Toolkit, MySQL, and other programs.

This is a standard “check for updates automatically” feature, with two additional features. First, the tool checks its own version and also the versions of the following software: operating system, Percona Monitoring and Management (PMM), MySQL, Perl, MySQL driver for Perl (DBD::mysql), and Percona Toolkit. Second, it checks for and warns about versions with known problems. For example, MySQL 5.5.25 had a critical bug and was re-released as 5.5.25a.

A secure connection to Percona’s Version Check database server is done to perform these checks. Each request is logged by the server, including software version numbers and unique ID of the checked system. The ID is generated by the Percona Toolkit installation script or when the Version Check database call is done for the first time.

Any updates or known problems are printed to STDOUT before the tool’s normal output. This feature should never interfere with the normal operation of the tool.

For more information, visit https://www.percona.com/doc/percona-toolkit/LATEST/version-check.html.

ENVIRONMENT

The environment variable PTDEBUG enables verbose debugging output to STDERR. To enable debugging and capture all output to a file, run the tool like:

PTDEBUG=1 pt-diskstats ... > FILE 2>&1

Be careful: debugging output is voluminous and can generate several megabytes of output.

ATTENTION

Using <PTDEBUG> might expose passwords. When debug is enabled, all command line parameters are shown in the output.

SYSTEM REQUIREMENTS

This tool requires Perl v5.8.0 or newer and the /proc filesystem, unless reading from files.

BUGS

For a list of known bugs, see https://jira.percona.com/projects/PT/issues.

Please report bugs at https://jira.percona.com/projects/PT. Include the following information in your bug report:

  • Complete command-line used to run the tool

  • Tool --version

  • MySQL version of all servers involved

  • Output from the tool including STDERR

  • Input files (log/dump/config files, etc.)

If possible, include debugging output by running the tool with PTDEBUG; see “ENVIRONMENT”.

DOWNLOADING

Visit http://www.percona.com/software/percona-toolkit/ to download the latest release of Percona Toolkit. Or, get the latest release from the command line:

wget percona.com/get/percona-toolkit.tar.gz

wget percona.com/get/percona-toolkit.rpm

wget percona.com/get/percona-toolkit.deb

You can also get individual tools from the latest release:

wget percona.com/get/TOOL

Replace TOOL with the name of any tool.

AUTHORS

Baron Schwartz, Brian Fraser, and Daniel Nichter

ABOUT PERCONA TOOLKIT

This tool is part of Percona Toolkit, a collection of advanced command-line tools for MySQL developed by Percona. Percona Toolkit was forked from two projects in June, 2011: Maatkit and Aspersa. Those projects were created by Baron Schwartz and primarily developed by him and Daniel Nichter. Visit http://www.percona.com/software/ to learn about other free, open-source software from Percona.

VERSION

pt-diskstats 3.6.0